RoamingPigs/trcs-tactical-audio

TRCS: a synthetic radio-channel benchmark for speech recognition

TRCS (Tactical Radio Channel Simulator) is 1,050 short audio files built to test one failure: speech models writing words that were never spoken. It also measures how word error grows when speech passes through telephone codecs and noise.

The speech is synthetic. Scripted radio calls are rendered with a text-to-speech engine (espeak-ng, US English, 155 words per minute) and then degraded in software. The codecs are real encoders and decoders. The noise and squelch effects are my own models. Nothing here was recorded from a radio, and the data does not represent any agency's traffic.

Self-published and not peer reviewed. Written by Cisco Caceres. Results, method and corrections are on roamingpigs.org. The evaluation code is at github.com/RoamingPigs/trcs.

Contents

Split	Files	What is in it
hallucination	90	9 conditions, 10 seeds each, 2.5 s clips: digital silence, quiet room, radio squelch burst, white, pink, brown, engine and babble noise, plus one speech control.
telephone_grid	960	40 scripted radio calls, 6 audio paths (uncompressed, G.711 mu-law, G.711 A-law, AMR-NB 12.2 kbit/s, GSM full-rate, Opus narrowband), 4 noise levels (clean, 20, 10 and 0 dB).

All files are 16 kHz mono 16-bit WAV. Total 2.27 hours and 261 MB.

How the audio was made

1. Speech: espeak-ng, voice en-us, 155 words per minute, resampled to 16 kHz.
2. Band-pass and codec: every narrowband path first resamples to 8 kHz and applies a 300 to 3400 Hz Butterworth band-pass, then encodes, decodes and returns to 16 kHz. G.711 uses the Python audioop companding functions. GSM full-rate and Opus (12 kbit/s, voice mode) use ffmpeg with libgsm and libopus. AMR-NB (12.2 kbit/s) uses the OpenCORE library. A result for a narrowband path therefore combines band-limiting and coding. The generator is at github.com/RoamingPigs/trcs in generator/. Regenerating all 1,050 files with it reproduced every SHA-256 in provenance_manifest.json on 2026-10-12. This paragraph corrected an earlier version of this card that said G.711 used ffmpeg and did not mention the band-pass. The audio files did not change.
3. Noise: pink noise is mixed in after the codec. The signal-to-noise ratio is measured over the whole clip including pauses, not over active speech only.

Verify the files

provenance_manifest.json lists a SHA-256 for every file. This re-hashes them all:

import hashlib, json, pathlib
root = pathlib.Path(".")
m = json.loads((root / "provenance_manifest.json").read_text())
bad = [f["filename"] for f in m["files"]
       if hashlib.sha256((root / f["relative_path"]).read_bytes()).hexdigest() != f["sha256"]]
print(len(m["files"]), "files,", len(bad), "mismatches")

A matching hash proves the file is the one listed. It does not prove the audio is realistic. There is no combined root digest, only one digest per file.

Known limits

• Synthetic speech. Real radio speech is faster, more varied and often clipped. Results may not carry over.
• Simulated channel. Noise is added after the codec, whereas on a real link noise arrives before it. SNR is defined over the whole clip.
• Duplicates. 18 files are byte-identical to another file (the digital-silence and speech-control clips do not vary by seed), so there are 1,032 distinct digests.
• 40 scripts. The scripted calls recur in every audio path and noise level, so results should be resampled by script, not by clip.
• One author. Nobody else has re-run the benchmark yet.

License

Audio fixtures and metadata: CC BY 4.0. Code in this repository: Apache-2.0. The dataset card and results pages say which version of the files a result used.

Cite

No DOI exists yet. See roamingpigs.org/cite for BibTeX. Mistakes are logged at roamingpigs.org/corrections.